All terms
Glossary
Reranking
A second step in retrieval: the first hits are reordered by a more accurate model before they reach the language model.
Fast vector search returns twenty roughly fitting sections. A reranker scores those twenty more precisely and passes on the best five.
It costs little, because only a small set is scored precisely, and it often lifts hit quality noticeably.
How you notice it
- The right answer is among the hits but far down.
- Many chunks are passed in and quality does not improve.
- Cost per request should come down.
Frequently asked
Is the extra step worth it?
Almost always. Retrieve twenty candidates roughly, rank the best five properly and pass only those: it improves the answer and lowers cost at the same time, because less text reaches the model.